Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Game Generation via Large Language Models

Chengpeng Hu, Yunlong Zhao, Jialin Liu

TL;DR

The paper examines generating games with large language models by jointly producing game rules and levels in Video Game Description Language (VGDL). It introduces the LLMGG framework, which prompts LLMs to generate VGDL representations parsed by GVGAI engines, and systematically studies prompt design through maze-generation experiments across GPT-3.5, GPT-4, and Gemma 7B. Key findings show that including explicit context substantially improves output parsability and correctness, with GPT-4 achieving perfect results in the best prompt configuration (P7) for all trials, while hallucinations and misinterpretations of VGDL semantics remain major challenges. The work highlights the potential for accessible, prompt-driven game prototyping, underscores the need for robust validation and human-in-the-loop intervention, and points to future work extending to 3D game generation and more advanced prompting strategies.

Abstract

Recently, the emergence of large language models (LLMs) has unlocked new opportunities for procedural content generation. However, recent attempts mainly focus on level generation for specific games with defined game rules such as Super Mario Bros. and Zelda. This paper investigates the game generation via LLMs. Based on video game description language, this paper proposes an LLM-based framework to generate game rules and levels simultaneously. Experiments demonstrate how the framework works with prompts considering different combinations of context. Our findings extend the current applications of LLMs and offer new insights for generating new games in the area of procedural content generation.

Game Generation via Large Language Models

TL;DR

The paper examines generating games with large language models by jointly producing game rules and levels in Video Game Description Language (VGDL). It introduces the LLMGG framework, which prompts LLMs to generate VGDL representations parsed by GVGAI engines, and systematically studies prompt design through maze-generation experiments across GPT-3.5, GPT-4, and Gemma 7B. Key findings show that including explicit context substantially improves output parsability and correctness, with GPT-4 achieving perfect results in the best prompt configuration (P7) for all trials, while hallucinations and misinterpretations of VGDL semantics remain major challenges. The work highlights the potential for accessible, prompt-driven game prototyping, underscores the need for robust validation and human-in-the-loop intervention, and points to future work extending to 3D game generation and more advanced prompting strategies.

Abstract

Recently, the emergence of large language models (LLMs) has unlocked new opportunities for procedural content generation. However, recent attempts mainly focus on level generation for specific games with defined game rules such as Super Mario Bros. and Zelda. This paper investigates the game generation via LLMs. Based on video game description language, this paper proposes an LLM-based framework to generate game rules and levels simultaneously. Experiments demonstrate how the framework works with prompts considering different combinations of context. Our findings extend the current applications of LLMs and offer new insights for generating new games in the area of procedural content generation.
Paper Structure (49 sections, 2 figures, 3 tables)

This paper contains 49 sections, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Preliminary
  3. LLMs and Procedural Content Generation
  4. Video Game Description Language
  5. Generating Games via LLMs
  6. Workflow of LLMGG
  7. Prompt design
  8. Generating Maze via LLMs
  9. Practical prompts for Maze
  10. Rule-based text validation
  11. Necessity of context
  12. Wrong game logic due to LLM hallucination
  13. Hard to self-correct
  14. Avoiding hallucination by following LLMs?
  15. Conclusion
  16. ...and 34 more sections

Figures (2)

  • Figure 1: Framework of game generation via LLMs (LLMGG). An LLM receives a prompt to generate game rules and levels in the representation of VGDL.
  • Figure 2: An example prompt of $P_6$, where "$\cdots$" denotes omitted texts.